Abstract
In the rapidly evolving domain of the Artificial Intelligence of Things (AIoT), ensuring effective person recognition, device authentication, and secure communications is becoming ever more critical. However, existing gait recognition systems typically demand extensive, environment-specific Radio Frequency (RF) data, making them difficult to scale. Widely adopted authentication schemes (e.g., QR codes, one-time passwords) offer only one-way verification and remain vulnerable to impersonation or replay attacks—especially threatening in scenarios such as unmanned vehicle deliveries. Meanwhile, connected vehicles in the Internet of Vehicles (IoV) ecosystem face high mobility and ad-hoc connectivity challenges, where most established key generation methods are ill-suited for long-range, low-power communication technologies like LoRa. Consequently, these combined issues hinder the practical deployment of robust AIoT services in real-world environments.In this dissertation, we systematically explore three core areas of AIoT security and sensing to address existing gaps in data collection, authentication robustness, and secure communications:
Firstly, we tackle the challenge of recognizing individuals through gait, a distinctive biometric trait, without the burden of extensive environment-specific data collection. This effort leads to a cross-modal framework, termed XGait, which simulates RF signals from IMU data readily available in mobile devices. By proposing an RF spectrogram generation method to consistently extract core features, applying a generative network to map IMU data into RF representations, and designing a dedicated transformer model for spectrogram analysis, we achieve over 99% Top-3 accuracy across multiple environments, RF devices, and mobile devices.
Furthermore, we focus on enhancing device authentication in scenarios like unmanned vehicle delivery, where high-value packages must be safeguarded against unauthorized interception. Our solution—Wave-for-Safe (W4S)—introduces mutual authentication by correlating random hand-waving gestures captured via multi-modal sensors on both the user’s smartphone and the unmanned vehicle (e.g., mmWave radar, camera). By carefully handling heterogeneous data processing, synchronizing sensor streams, and mitigating imitation attacks, W4S consistently delivers an equal error rate below 0.013, even in diverse real-world tests.
Finally, we address the security bottleneck in IoV (Internet of Vehicles) communications, where high mobility and ad-hoc connectivity demand more robust approaches to secret key establishment. To that end, we propose Vehicle-Key, a physical-layer key generation system tailored for the long-range, low-power LoRa communication technology. This design employs a deep learning model capable of simultaneous channel prediction and quantization, complemented by an autoencoder-based reconciliation process that significantly boosts key agreement rates. Real-world experiments indicate 15.10%–49.81% higher key agreement and a 9–14× increase in key generation speed, with a minimal runtime of 3.4 ms on a Raspberry Pi—demonstrating both efficiency and strong security guarantees.
Overall, this dissertation effectively bridges critical gaps in AIoT ecosystems. The cross-modal gait recognition framework establishes accurate person recognition, the mutual authentication system safeguards high-value deliveries, and the physical layer approach secures vehicular communications at scale. Collectively, they form a comprehensive foundation for smart, resilient, and trustworthy everyday AIoT applications.
| Date of Award | 11 Jun 2025 |
|---|---|
| Original language | English |
| Awarding Institution |
|
| Supervisor | Weitao XU (Supervisor) |